Comparison of the performance of vector quantiser training algorithms

The vector quantiser (VQ) is a single layer winner-takes-all neural network. The `transformation´ performed by the VQ consists of two components, an encoding operation and a code book that stores the allowed codes and the vectors that correspond to them, known as code vectors. The VQ can also be used to perform lossy data compression, but it is not robust to corruption of the code indices. The Kohonen self-organising map is a close relative of the VQ, and would appear to be a good candidate to provide encodings that are robust to channel noise, because of the topological ordering of the code vectors. However Luttrell has shown that a kind of VQ, known as the topographic VQ (TVQ) emerges naturally when considering a single VQ transmitting codes along a noisy communications channel. The theory regarding the training and encoding TVQ is reviewed within the context of a noisy communications channel. Simulations are presented showing that TVQ´s are more robust to noise than the Kohonen map. Lastly, it is shown that the TVQ has the same code vector density as the VQ